Mira Murati's Thinking Machines drops Inkling, an open-weight AI that knows when it's guessing

Mira Murati’s new startup just showed its cards, and they’re not what you’d expect from a former OpenAI CTO. Thinking Machines Lab released Inkling on Wednesday, a 975-billion-parameter mixture-of-experts model that’s completely open-weight. Anyone can download it, tweak it, and run it on their own terms — a direct shot across the bow of the locked-down, API-walled gardens that dominate the market.

The model is massive but efficient, activating only about 41 billion parameters per task. It was trained on a hefty 45 trillion tokens spanning text, images, audio, and video, and reasons across all those modalities natively. But raw power isn’t the pitch. The company’s own briefing materials are blunt: Inkling is “not the strongest model available today, closed or open.” What it does differently is admit when it doesn’t know something. It’s designed to flag uncertainty rather than hallucinate confidently, and users can dial a “thinking effort” knob up or down to trade speed for depth.

This release is the culmination of a bet that enterprises don’t want one-size-fits-all AI. A recent blog post from the company laid out the thesis plainly: centrally trained models underperform because they can’t absorb the specific expertise locked inside an organization. Microsoft CEO Satya Nadella echoed a similar warning just days ago, arguing that companies using proprietary models pay twice — once in fees, and again by handing over proprietary business knowledge that gets baked into future versions of the model. Hugging Face CEO Clem Delangue told TechCrunch he sees most production AI work shifting to private or open-source alternatives.

The most compelling evidence Thinking Machines offers is a project with Bridgewater Associates. The two took an existing open-source model and fine-tuned it on the hedge fund’s internal financial expertise. The resulting system reportedly scored 84.7% on financial reasoning tests, beating top proprietary models while costing roughly a fourteenth as much to run. Those results are self-reported, not independently verified, so take them with a grain of salt. But the speed is real: Thinking Machines says it went from founding to a shipping product in about nine months, a fraction of the time it took OpenAI or Anthropic. On the uncomfortable question of whether Inkling was trained on outputs from competitors’ models, the company concedes it partly was — a practice known as distillation that remains a legal and ethical gray area across the industry.